Harness²

Harness² treats the harness as eight editable pieces — system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins — while the model stays frozen. Task-level recursion runs Propose–Probe–Compose: propose harness edits, probe them on the same task, and compose a refined harness from behavioral contrasts, without ground-truth feedback. Harnessing-level recursion then keeps adopted harnesses and evidence-backed lessons so the next task starts smarter. Executors include OpenCode and Codex; benchmarks cover JobBench, LAB, and WorkBuddy-Bench.

Most teams still ship a fixed agent scaffold and hope the model is smart enough. Real work fails on open tasks with no grader, so you need a way to rewrite the harness from behavior, not labels. Harness² is the research-grade loop for that: better agents by editing the system around the model, with measured gains on professional benchmarks.

On Linux with Python 3.12 and uv, install with uv pip install -e '.[render]', then harness2-demo for a no-credentials dry run. For real runs, set Vertex or other models in .env, run harness2-setup for a bench, pick --k and --mode parallel or sequential, then harness2, harness2-judge, and harness2-report. Scale k for wider parallel probes or deeper sequential refinement. Results land under runs/ and resume if you repeat the same command.

Not another coding agent TUI and not a skill pack dump. Harness² is recursive optimization of the scaffold itself — contrastive edits without a grader, then a persistent improvement harness across tasks. Complementary to AutoHarness (distill sessions into skills), ECC (agent harness OS), and harness-engineering (org field guides): those equip or run agents; this one systematically rewrites the harness and proves it on open-world professional work.

• Public demo available

• AutoHarness ↗
• ECC ↗
• Harness Engineering ↗